Translating Product Descriptions: AI, but How?

A French store that wants to sell in Germany, Spain, and the Netherlands can triple its potential market. The price to pay: translating 5,000 product pages into 7 languages, or 35,000 translations. At €20 per page with a freelance translator, that means €700,000 and 18 months of lead time — unrealistic for 99% of stores.

In 2026, AI-targeted machine translation makes this project possible in 2–3 weeks for under €1,000. But between "doing it" and "doing it well," there are 4 pitfalls to avoid.

Pitfall 1: DeepL or Google Translate alone are not enough

Generic machine translation engines (DeepL, Google, Microsoft) are excellent for standard prose. They fail in e-commerce for three reasons:

Product vocabulary is market-specific. In British English, pants are called "trousers," while in American English they are "pants" — and "pants" in the UK means underwear. An uncalibrated machine translator does not know that.

Local units and formats. FR size 44 = US 10 = UK 10 = CN 280. A 75 cm high chair in France becomes 29.5" in the US. A 5000 mAh battery stays 5000 mAh everywhere — but "2 days of battery life" may become "2 days" or "48 hours" depending on the market's commercial tone.

Cultural references. "Ideal for pre-dinner drinks with friends" does not translate word for word into German. In Spain, "tapas" may replace "appetizers." These adaptations break a generic translator.

The solution: a translation pipeline with an e-commerce-specific brief, plus post-processing that detects and corrects local errors.

Pitfall 2: slugs and URLs that should be localized (or not)

When you translate a blog post or a product page, the URL question comes up.

Product slug: should it be translated?

✅ For local SEO, yes. An EN slug like leather-derby-shoes ranks better in the US than an FR slug like chaussure-derby-cuir.
❌ For maintenance, it is costly — 5,000 products × 7 languages = 35,000 slugs to manage, with redirects when changes happen.

The right approach: use hreflang infrastructure to tell Google which language/market matches which URL, and use localized slugs by language (not by country):

/fr/products/chaussure-derby-cuir
/en/products/leather-derby-shoes
/de/products/derby-lederschuhe
/es/products/zapatos-derby-cuero

One slug per language, not one per country. Regional variants (fr-CA vs fr-FR, en-US vs en-GB) should be handled through hreflang and server-side geographic targeting, not through separate URLs.

See our full guide on multilingual e-commerce hreflang.

Pitfall 3: special characters and encoding

A real-world problem: an FR store expanding into German sees its "ß" become "?" in product titles. Its meta description "für Damen" shows up as "fr Damen." Its URLs contain unreadable URL-encoded characters.

Cause: the translation pipeline or database is not configured in UTF-8. In 2026, this is rare — but not nonexistent, especially on older WooCommerce setups (MySQL in latin1 by default).

Checklist:

Database in utf8mb4 (not plain utf8 — utf8 does not handle all emojis and some Asian characters)
Product table collation in utf8mb4_unicode_ci or utf8mb4_0900_ai_ci
HTTP headers: Content-Type: text/html; charset=utf-8
Slugs without accents or special characters (no ü, ß, ñ in the URL — use ASCII equivalents or transliterations)

Shopify handles UTF-8 natively without configuration. WooCommerce requires checking wp-config.php and the host-side MySQL configuration.

Pitfall 4: human post-editing you do not have time to do

The ideal plan: AI translates → human reviews → publish. In reality, across 35,000 translations, nobody reviews all 35,000. Here is the realistic approach:

Tiered strategy by ROI

Top 20% of the catalog by revenue → systematic human review (native speaker from the target market). Typically 200–1000 pages.
Middle 60% → sample-based review (10% of pages). This catches systemic errors without scaling costs.
Bottom 20% → AI only + automatic anomaly detection (abnormal length, unreadable characters, too many source-language words).

This strategy reduces the human review workload to ~15% of the catalog while covering 85% of revenue. It is the trade-off that makes multilingual e-commerce realistic.

The Ecomptimize method in 4 steps

Here is the pipeline we use to translate catalogs of 5,000 to 50,000 pages:

Step 1 — Create the market brief

For each target language, define:

The tone (formal vs informal)
The specific vocabulary (lists of terms not to translate, required translations for certain words)
Units and formats (sizes, weight, currency)
Cultural references to adapt (yes/no)

This brief is created once per language, in 1–2 hours. It is reused for all future translations.

Step 2 — Targeted AI translation

GPT-5.4 (or Claude Opus) receives the source FR page + the target market brief. It produces:

The translated title calibrated for local SEO
The full description (benefits rewritten, not just translated)
The meta description (re-optimized for CTR in the target market, not a direct translation)
The translated tags
A candidate slug (or 3 candidates to validate)

Step 3 — Automatic validation

A script checks:

No source-language words left over ("le," "de," "the," "the" in the EN version...)
Length within limits (title < 60, meta < 160)
Correct special characters for the language
Valid slug (no forbidden characters)

Pages that fail are sent to a manual review queue.

Step 4 — Batch publishing

Validated pages are published in bulk through the admin API (via POST /admin-cms/blog/articles with locale=xx, or the product equivalent). Rollback is possible if something goes wrong.

Total cost for 7 languages × 5,000 pages

API cost (GPT-5.4 premium): ~€400–600 for 35,000 translations
Orchestration tool (Ecomptimize or equivalent): included in the subscription
Human post-editing (20% of the catalog): ~€3,000–5,000 with multilingual freelancers
Total: €3,500 to €5,600 to go from one language to 7 languages

Compare that with €700,000 for a traditional 100% human translation workflow. Ratio: 125× cheaper.

FAQ

How many languages should an EU e-commerce store support?

To cover 80% of the EU market: FR, EN, ES, DE, IT. These 5 languages represent ~350 million consumers. Adding PT and NL brings coverage to ~90%. Beyond that, ROI drops quickly — prioritize depth (quality in the first 5) over expansion (adding Polish, Czech, etc.).

Should customer reviews be translated?

No. Reviews in the source language signal authenticity. Google does not penalize a mix of languages in reviews — it can even be a trust factor. Exception: if 90% of your reviews are in one language and you are expanding internationally, add subtitles or translations for the main reviews at the bottom of the page.

Should I automatically detect the visitor's language?

By IP, no (VPNs break detection and create traps for Google crawlers). By the browser's Accept-Language header, yes, but carefully. The right 2026 approach: explicit localized URLs (/fr/..., /de/...) + a persistent language selector visible in the header.

How do I handle regional variants of the same language (fr-FR vs fr-CA)?

Through hreflang in the sitemap and in <link rel="alternate"> tags. Content can be identical for 90% of pages, with local variants (price, currency, expressions) for the top 10% of the catalog. Search engines serve the right version based on the visitor's IP geolocation.

Do AI-translated descriptions meet Google's EEAT criteria?

Yes, under one condition. EEAT (Experience, Expertise, Authoritativeness, Trustworthiness) measures the overall reliability of your site, not each description in isolation. A catalog with AI translations + accurate product information + clear policies + authentic reviews is fine. A catalog with AI translations + thin content + missing legal information will be penalized — but the problem is not AI, it is overall quality.

How long does it take to launch a new language on a 5,000-page catalog?

With the AI pipeline: 3–5 days total, including brief creation, translation, validation, and publishing. With human post-editing on 20% of the catalog: add 2–3 weeks for review. The first language takes the most time (pipeline setup); the next ones can be deployed in 2–3 days each.

To launch your catalog in 7 languages in 2–3 weeks, see Ecomptimize for Shopify or Ecomptimize for WooCommerce.